From Signal to Context: How We Feed Our Agents

There is a difference between giving an agent a number and giving it a number with meaning. For a while, our agents got the number. Here is what changed — and what that looks like on a real signal.

A real signal, traced through the full pipeline

On May 22, 2026, FIIs sold ₹4,440 Cr in the cash market. DIIs bought ₹6,003 Cr. This is what that signal looked like at each stage of the pipeline — from raw number to agent context.

Signal trace · May 22, 2026 · FII/DII flows

Stage 1 — Raw sensor output

What fii_dii_scraper.py returned from the NSE data feed:

FII_cash_net: -4440.0 Cr
DII_cash_net: +6003.0 Cr
date: 2026-05-22
session: cash_market

Stage 2 — Domain-scoped interpretation

The same data goes through five nodes, each with a different lens. Market Oracle — mandate: analyze institutional flows — produced:

BEARISH | FII net: −₹4,440 Cr cash market | Institutional exit accelerating; DII absorption strong but selling pressure is sustained not opportunistic

Macro Intelligence — mandate: monitor global risks — connected it to the rate environment:

NEUTRAL | Domestic liquidity index: elevated post-RBI cut | FII exit cushioned by domestic liquidity; not a systemic signal at current pace

Stage 3 — Sentinel verification

The Sentinel queried Tavily with a status-aware query — markets were live at time of sync. Both outputs verified against current NSE data. Both marked sentinel_verified: true.

The graph simultaneously built the causal chain from these verified outputs:

FIIs → TRIGGERS → Foreign Capitulation (w=0.85)
Foreign Capitulation → TRIGGERS_BUYING → DIIs (w=0.85)
DIIs → SUPPORTS → Indian Market (w=0.85)
Domestic Liquidity → CUSHIONS_POSITIVELY → Foreign Capitulation (w=0.85)

Stage 4 — What a sovereign agent received

A sovereign agent monitoring Indian large-cap exposure received this as its context input — not the raw numbers from Stage 1:

[MACRO_PILLAR] NEUTRAL | Domestic liquidity elevated | FII exit cushioned by rate environment
[INSTITUTIONAL] BEARISH | FII net −₹4,440 Cr | Sustained exit, not opportunistic
[SYNTHESIS] Foreign capital exiting; domestic institutions absorbing more than the full outflow; market holding. Watch for three-session continuation to confirm trend vs. rebalancing.

The raw number was -4440.0. What the agent reasoned from was a directional bias, a domain-specific metric, a causal chain, and a synthesis — all sentinel-verified before the agent saw any of it.

The pipeline that produced this

Signal fetch

Raw readings across active sensors. Prices, flows, volumes. Uninterpreted.

Domain layering

5 nodes. Same data, different lens. Domain boundary enforced per node.

Sentinel audit

Tavily verifies every output. Status-aware. Nothing passes unchecked.

Agent feed

Verified, interpreted context goes to hired and sovereign agents.

The five domain nodes

Each node receives the same underlying market data. Each has a domain boundary enforced in its system prompt — not as a guideline, as a constraint. Market Oracle is not asked to have a view on sentiment. Sentiment Radar is not asked to interpret yield curves.

Node	Domain	Mandate
Sentiment Radar	sentiment	Track retail sentiment
Market Oracle	institutional	Analyze dark pool flows
Macro Intelligence	macro	Monitor global risks
Bharat News Agent	corporate	Track Indian catalysts
Neural Briefing	synthesis	Synthesize all signals

Enforced output format on every node

BIAS — BULLISH / BEARISH / NEUTRAL

METRIC — specific data point from that node's domain only

INSIGHT — one sentence

An agent that can say anything says nothing useful. The format forces specificity. METRIC must be domain-specific — not a generic market observation.

The Sentinel — what it actually caught

Before any output is written, a grounding layer runs. On the same day the FII/DII signal above was verified correctly, the Sentinel also caught this:

Sentinel flagged · May 23, 2026

ClaimedSalil Parekh made a market purchase of INFY shares worth ₹420.5 lakhs in mid-May 2026 — graph edge: INSIDER_BUY, bullish signal.

RealitySEBI filings: he exercised grants on ~May 2, then sold 95,800 shares on May 8. No open-market purchase. Edge removed. Signal was inverted.

The status-aware logic also works in the other direction. If Neural Briefing says "market closed at X" during live trading hours, the Sentinel corrects it before anything is written downstream.

Agent output

Unverified. May reflect stale data, status errors, or wrong source.

→

Sentinel

Status-aware Tavily query. Live market vs. closed changes the query type entirely.

→

Result

sentinel_verified: true — or flagged, corrected, and quarantined before it reaches any agent.

What changed for agent output

Before

Agents received raw sensor readings. A number and a timestamp. Output quality was bounded by how much reasoning an agent could do with an uninterpreted number.

Now

Agents receive domain-interpreted, sentinel-verified context. Neural Briefing synthesizes across all five domain nodes before downstream agents see anything.

The depth of the input determines the ceiling of the output. A hired agent reasoning about a specific position isn't starting from -4440.0. It's starting from a verified directional bias, a domain-specific metric, and a causal chain — all assembled before it runs.

What a sovereign agent receives today

A sovereign agent monitoring Indian large-cap exposure currently receives:

✓

Macro Intelligence — NEUTRAL · domestic liquidity elevated post-RBI cut · FII exit cushioned by rate environment · sentinel-verified

✓

Market Oracle — BEARISH · FII net −₹4,440 Cr cash market · sustained institutional exit, not opportunistic rotation · sentinel-verified

✓

Bharat News Agent — reading on Indian corporate catalysts for the session · sentinel-verified

✓

Neural Briefing — synthesis: foreign exit absorbed by domestic institutions; market holding; watch for continuation signal · sentinel-verified

It does not receive a raw NIFTY price and a timestamp.

The next layer: graph context

The pipeline today handles depth on one axis — domain-scoped interpretation of each signal. The context graph, once clean, adds the second axis: what each signal connects to across domains.

How the two layers connect

Domain-scoped interpretation per node — working today. The FII/DII signal above is a live example.

Sentinel verification before every write — working today. Caught the inverted INFY insider signal before it reached any agent.

Entity resolution + controlled vocabulary in the graph — in progress. Required before multi-hop traversal is reliable.

Graph traversal feeding causal chains into agent context — next. Yield curve → FII behavior → sector rotation → specific company exposure, in one path, traceable at every hop.

Next: what the graph looks like after seven days of 27-sensor operation, and how it begins feeding this pipeline.